FAQ#
Answers to questions that come up repeatedly in the issue tracker and on Discord. If you don’t see your question here, check Getting help for where to ask.
Getting started#
What hardware do I need to run FlashDreams?#
FlashDreams targets recent NVIDIA data-center GPUs. The profiling corpus on the Models page runs across three devices — NVIDIA H100, GB200, and GB300 — and the per-recipe latency numbers on each model page (e.g. Self-Forcing, LingBot-World) are sourced from the same set.
Any CUDA-capable GPU with enough memory for the chosen checkpoint should run the streaming recipes; smaller GPUs may need to drop multi-GPU recipes back to a single device. See the Get Started for the cheapest path to a working clip, and the Models page for the profiled configurations.
Which model recipes ship in the box?#
First-party model integrations ship under integrations/ in the
repo. The Models page has the full list; the documented
recipes are:
NVIDIA OmniDreams — HDMap-conditioned driving world model (the GTC 2026 closed-loop demo).
Self-Forcing — Streaming Wan 2.1 T2V (1.3B).
Causal-Forcing — Streaming Wan 2.1 T2V / I2V (1.3B).
Causal Wan2.2 — FastVideo Causal Wan 2.2 14B MoE T2V.
LingBot-World — Camera-controllable I2V world model.
FlashVSR — Streaming video super-resolution.
Wan2.1 — Bidirectional Wan 2.1 T2V / I2V reference.
Cosmos-Predict2.5 — Bidirectional Cosmos-Predict2.5 T2V / I2V reference.
Each model page has the canonical CLI invocation, checkpoint source, multi-GPU command, and per-recipe knobs.
Installation and packaging#
Why can I install flashdreams from PyPI but not the integration packages?#
Only the core flashdreams package is published as a pure-Python
wheel on PyPI. Integration packages — flashdreams-self-forcing,
flashdreams-lingbot, and the others listed in
DEV.md —
are not published; they live under integrations/ in the monorepo
and are designed to be consumed either as a workspace member or as
git-installable packages.
To install an integration directly from the repo:
pip install "flashdreams-wan21 @ git+https://github.com/NVIDIA/flashdreams.git#subdirectory=integrations/wan21"
# or with uv
uv pip install "flashdreams-wan21 @ git+https://github.com/NVIDIA/flashdreams.git#subdirectory=integrations/wan21"
The rationale is in DEV.md: the
core flashdreams is the only stable, pip-installable surface; the
per-recipe wheels move at the upstream model’s pace and stay
git-installable so they can pin against a known core commit.
Usage#
How do I plug in a new model recipe?#
The Add a new method guide walks the full
flow — what to subclass on the runner side, how the entry-point
registration works, and what the per-integration directory layout
looks like. The in-tree integrations under integrations/
are the canonical references; pick the one closest in shape to your
new recipe and use it as a template.
The minimum surface is: subclass the right runner base, register the
slug via the flashdreams.runner_configs entry point, and (for
streaming runners) wire --total-blocks into the runner config.
Add a new method covers each step with the
exact pyproject.toml snippet.
Project and licensing#
Can I use FlashDreams commercially?#
Yes. FlashDreams is released under the Apache License 2.0, which permits commercial use, modification, and distribution under the license’s terms. Third-party model weights and datasets used with FlashDreams may carry their own licenses — please check those separately.
Contributing back is welcome but not required. See Contributing to FlashDreams if you’d like to upstream a fix or improvement.